COMPARISON OF DEEP LEARNING ALGORITHMS FOR LEUKEMIA CANCER CELL CLASSIFICATION

Abstract

Leukemia is a cancer-related disease which causes the death of individuals worldwide, regardless of age and gender. It affects the blood and bone marrow, thus leading to the abnormal production of immature white blood cells. Some of the factors that might contribute to leukemia’s development might be related to genetics, radiation or chemical exposure, infections, or immune system disorders. A reliable and fast diagnosis of leukemia is crucial for a successful treatment to ensure high survival rates and low number of deaths. Nowadays, blood tests are widely used for diagnosing leukemia. Patients undergo a complete blood count (CBC) to evaluate the count of blood cells present. In cases of leukemia, CBC reveals abnormal count of white blood cells (WBC), red blood cells (RBC) and platelets. Additionally, these blood cells are examined under a microscope. Based on the results, immature or abnormal-looking white blood cells may indicate leukemia. However, this type of diagnosis is often slow, time-consuming and less accurate, mainly because under microscopes, the shape of leukemic cells might seem similar to the shape of normal white cells, therefore making the diagnosis prone to errors. Therefore, in this thesis, we will focus on the deep learning algorithms which have shown promising results in diagnosing leukemia cells. Some of these algorithms include Convolutional Neural Networks (CNNs), which in the context of leukemia cells diagnosis, can be trained to classify images of blood smears into normal blood cells or leukemic blood cells. The second algorithm includes Optimized Deep Recurrent Neural Networks (ODRNNs), which can be used to analyze time-series data such as videos of cell movements or changes in cell morphology over time. The last algorithm is Transfer Learning, which is applied by fine-tuning a pre-trained neural network on a dataset of leukemia cells. This approach helps improve the performance of the model, especially when limited labelled data are available for training.